• No se han encontrado resultados

Línies bàsiques de treball

In document Vista de Exemplar complet (página 55-60)

Los inicios del aprendizaje-servicio en Catalunya

5. Línies bàsiques de treball

Then, we have

Vex(t, η) =

where In is defined in (2.5.11). Hence,

|res0| ≤ σ(t)Y

also provides the time-frequency zone for eVxk(t, η). To describe those time-frequency zones mathematically, let 0 < τ0 < 1 be a given small number and the threshold:

Ok = {(t, η) : |Gk

σ(t)(η − φ0k(t)

| > τ0, t ∈ R}. (2.6.6)

Assuming |Gk(ξ)| is even and decreasing for ξ ≥ 0, then we may write Ok as

Ok= {(t, η) : |η − φ0k(t)| < αk

σ(t), t ∈ R} (2.6.7) where αk = αk(t) is obtained by solving |Gk(ξ)| = τ0.In this instance, we will say that multicomponent signal x(t) is well-separated and there is σ(t) such that

Ok∩ Ol = ∅, k 6= l.

We can use a Gaussian function defined in (2.5.24) as an example. For this g, we can obtain (see [15]),

Gk(u) = 1

p1 − i2πφ00k(t)σ2(t)e

2u2

1 + (2πφ00k(t)σ2(t))2(1+i2πφ

00

k(t)σ2(t)))

.

Thus,

|Gk(u)| = 1

(1 + (2πφ00k(t)σ2(t))2) 1 4 e

2

1 + (2πφ00k(t)σ2(t))2) u2.

The solution of |Gk(u) = τ0| ⇔ e

2u2

1 + (2πφ00k(t)σ2(t))2 = τ0((1+(2πφ00k(t)σ2(t))2) 1 4 Therefore, in this case, assume τ0((1 + (2πφ00k(t)σ2(t))2)

1 4 ≤ 1,

αk = q

1 + (2πφ00k(t)σ2(t))2) 1 2π

r 2ln(1

τ0) −1

2ln(1 + (2πφ00k(t)σ2(t))2)).

The main theorem on the 2nd-order adptive FSST can be written in more notations observe:

Gj,k(t, η) = Z

R

ei2π(φ0k(t)τ +12φ00k(t)τ2) τj

σ(t)j+1g( τ

σ(t))e−i2πητdτ. (2.6.8)

= F

eiπ(φ00k(t)τ2)τjg(τ )

σ(t)(η − φ0k(t)). (2.6.9)

Clearly,

G0,k(t, η) = Gk σ(η − φ0k(t)),

and, when j ≥ 1, we can see that

Gj,k = 1

(−i2π)jG(j)k σ(t)(η − φ0k(t)). (2.6.10) Let res1, res2, res00, and res01 be the residuals defined as res0 in (2.6.2) with g(τ ) replaced respectively by g1(τ ), g2(τ ), g0(τ ), and g3(τ ) = τ g0(τ ). We therefore have the estimates for these residuals that are similar to (2.6.4).

|res1| ≤ σ(t)Y

1

(t), |res2| ≤ σ(t)Y

2

(t), |res00| ≤ σ(t) fY

0(t), |res01| ≤ σ(t) fY

1(t),

where

Y

1

(t) = Kε1I2+ π

3I4σ2(t)

K

X

k=1

Ak(t), Y

2

(t) = Kε1I3+ π

3I5σ2(t)

K

X

k=1

Ak(t),

Yf

0(t) = Kε1Ie1+ π

3Ie3σ2(t)

K

X

k=1

Ak(t), Yf

1(t) = Kε1Ie2+ π

3Ie4σ2(t)

K

X

k=1

Ak(t),

with In, eIn defined in (2.5.11).

Denote:

Bk(t, η) =X

l6=k

xl(t)(φ0l(t) − φ0k(t))G0,l(t, η), Dk(t, η) =X

l6=k

xl(t)(φ00l(t) − φ00k(t))G1,l(t, η), Ek(t, η) =X

l6=k

xl(t)(φ0l(t) − φ0k(t))G1,l(t, η), Fk(t, η) =X

l6=k

xl(t)(φ00l(t) − φ00k(t))G2,l(t, η)

and

Res1 = i2πBk(t, η) + i2πσDk(t, η) + i2π(η − φ0k(t))res0− res00

σ(t) − i2πφ00k(t)σ(t)res(2.6.11)1, Res2 = 4π2σ(t)Ek(t, η) + 4π2σ2(t)Fk(t) (2.6.12)

+ i2πres0+ 4π2(η − φ0k(t))σ(t)res1+ i2πres01− 4π2φ00k(t)σ2(t)res(2.6.13)2.

Lemma 2.6.2. [19] Let Res1 be the quantity defined by (2.6.11). Then,

tVex(t, η) =



i2πφ0k(t) − σ0(t) σ(t)



Vex(t, η) + i2πφ00k(t)σ(t) eVxg1(t, η) − σ0(t)

σ(t)) eVxg3(t, η) + Res(2.6.14)1

Lemma 2.6.3. [19] For (t, η) satisfying eVx(t, η) 6= 0 and ∂

∂η(Vexg1(t, η Vex(t, η)) 6= 0, we have

P0(t, η) = i2πσφ00k(t) + Res3 (2.6.15)

where

Res3 = Vex(t, η)Res2− ∂ηVex(t, η)Res1

Vex(t, η)∂ηVexg1(t, η) − eVxg1(t, η)∂ηVex(t, η), (2.6.16) with Res1 and Res2 defined by (2.6.11)and (2.6.13) respectively.

Theorem 2.6.4. [19] Suppose x(t) ∈ D(2)ε12 with a window function g(t) for some small ε1, ε2 > 0. Then, we have the following:

(a) Supposeεe1 satisfiesεe1 ≥ ε0PK

k=1Ak(t) + σ(t)Π0(t). Then for (t, η) with

Vex(t, η)

>εe1, there exists k ∈ {1, 2, ..., K} such that (t, η) ∈ Ok. (b) Suppose (t, η) satisfies |Vx(t, η)| >εe1,

η( eVxg1(t, η)/ eVx(t, η))

>eε2, and (t, η) ∈ Ok. Then

ωapd,2nd,c(t, η) − φx0k(t) = Res4, (2.6.17)

where

Res4 = Res1

i2π eVx(t, η) − Vexg1(t, η)Res3

i2π eVx(t, η) . (2.6.18) Furthermore,

ωapd,2nd(t, η) − φ0k(t)

< Bd1 (2.6.19) where

Bd1 = max

1≤k≤K sup

η∈Ok

n|Res1|

2πeε1 + 1 2πεe31εe2

Vexg1(t, η)



ηVex(t, η)

|Res1| +εe1|Res2|o

(c) Suppose that ε1 satisfies the condition in part (a) and Bd1 ≤ 1 2Lk(t), where

Lk(t) = 1

σ(t)min{αk+ αk−1, αk+ αk+1}. (2.6.20) Then for any eε3 =εe3(t) > 0 satisfying Bd1 ≤eε3 ≤ 1

2Lk(t),

lim

λ→0

σ(t) g(0)

Z

|ξ−φ0k(t)|<eε3

Radp,2nd,λx,eε

1,eε2 (t, ξ)dξ − xk(t)

≤ Bd2, (2.6.21) where Bd2 = Bd02+ Bd002 with

Bd02 = 1

|g(0)|

n

k(εe1+ σ(t)Π0(t)) + Ak(t)

Z

|u|≥αk

Gk(u)du +X

l6=k

Al(t)Ml,k(t) o

,

Bd002 = 1

|g(0)|{2Π0(t)) + Ak(t) k g k1 |Zt| +X

l6=k

Al(t)Ml,k(t)}

and |Zt| represents the Lebesgue measure of the set Zt:

Zt=n

η : (t, η) ∈ Ok,

Vex(t, η)

>εe1, ,

∂(Vxg1(t, η)/ eVx(t, η))

≤εe2o .

Chapter 3

Higher-order SST

3.1 High-order Synchrosqueezing transform

3.1.1 The higher-order wavelet synchrosqueezing trans-form

Higher-order WSST

Consider the Nth-order polynomial-phase signal

x(t) = Aei2πφ(t) (3.1.1)

with φ(t) = r1t + 12r2t2+ · · · + N1rNtN, and φ0(t) = r1+ · · · + rNtN −1. We have x0(t) = i2πφ0(t)x(t) and

φ0(b + at) = φ0(b) +φ00(b)at

1! + · · · + φ(N )(b)aN −1tN −1

(N − 1)! . (3.1.2)

We recall the CWT definition of the signal x(t)

Wx(a, b) = Z

−∞

x(b + at)ψ(t)dt. (3.1.3)

Thus, we have

bWx(a, b) = Z

−∞

x0(b + at)ψ(t)dt = Z

−∞

i2πφ0(b + a)x(b + at)ψ(t)dt

= Z

−∞

i2π(r1+ r2(b + at) + · · · + rN(b + at)N −1)x(b + at)ψ(t)dt

= i2π Z

−∞



φ0(b) + φ00(b)at

1! + · · · + φ(N )(b)aN −1tN −1 (N − 1)!



.x(b + at)ψ(t)dt

= i2π

φ0(b)Wx(b, a) +φ(2)(b)a

1! Wxψ1(b, a) + · · · + aN −1(N −1)!φ(N )(b)WxψN −1(b, a) (3.1.4)

where

Wxψk(b, a) = Z

−∞

x(b + at)tkψ(t)dt. (3.1.5)

We can then write ∂bWx(b, a) =

N

X

k=1

i2πak−1

(k − 1)!φ(k)(b)Wxψk−1(b, a).

The goal is to determine φ(1)(b) according to Wx(b, a), Wxψ1(b, a), . . . , WxψN −1(b, a)

bWx(b, a)

i2πWx(b, a) = φ(1)(b) +

N

X

k=2

ak−1 (k − 1)!

Wxψk−1(b, a)

Wx(b, a) φ(k)(b). (3.1.6) We can write the equation in the form of a scalar product:

wx(a, b) = [x1(b), . . . , xN(b)][1, V2,1(b, a), . . . , VN,1(b, a)]T,

where wx(a, b) = ∂bWx(b, a)

i2πWx(b, a) and xk(b) = φ(k)(b), for k = 1, . . . , N.

To solve the problem, we pass through successive derivatives of equation

(3.1.6) according to the variable a.

To get sequence xk(b)

1≤k≤N, we create up a system of N equations with variables xk(b). Let us denote

y1 = VN.XNT (3.1.7)

with VN = [1, V2,1(b, a), . . . , VN,1(b, a)] and XN = [x1(b), . . . , xN(b)].

Computing the partial derivatives:

y2(b, a) = ∂ay1(b, a)

∂aV2,1(b, a) and Vk,2(b, a) = ∂aVk,1(b, a)

∂aV2,1(b, a), (3.1.8) which implies the following expression:

y2(b, a) = [0, 1, V3,2(b, a), . . . , VN,2(b, a)]XNT (3.1.9)

Vk,2(b, a) = ∂aVk,1(b, a)

∂aV2,1(b, a), for k = 3, . . . , N.

To get the jthequation. We repeat the same process iteratively. We define the new parameter for the AN matrix for j = 2, . . . , N and k = j, . . . , N by:

yj(b, a) = ∂ayj−1(b, a)

∂aVj,j−1(b, a), and Vk,j(b, a) = ∂aVk,j−1(b, a)

∂aVj,j−1(b, a). (3.1.10) Then,

yj(b, a) = [0, 0, . . . , 1, Vj+1,j, . . . , VN,j]XNT.

We group the N equations and get a good linear system

Since the AN is an upper triangular matrix with a nonzero diagonal, the solution of the linear system is given by

xN(b) = yN(b, a)

We can write this idea in the form of an algorithm.

Determination of the Nth-order local phase transformation

Step 1. We define the matrix AN by (3.1.12) with Vk,jobtained by the following formula:

AN = Step 3. We solve (3.1.11) and obtain

xN(b) = yN(b, a)

Step 4. The Nth-order phase transformation ωxN is defined by

ωxN(a, b) =

Adaptive higher-order WSST

For a signal with Nth-order polynomial-phase, we define adaptive higher order synchrosqueezing transform fωxN

. We start with the CWT with a time-varying parameter σ(t).

Wfx(a, b) =

The partial derivative of W (a, b) is given by

We can simplify the expression

bfWx(a, b) =

In addition for a signal x defined in (3.1.1), eβ(a, b) satisfies

Thus, we have

β(a, b)e

We can also put y1 in the form of a scalar product defined as follows

Similarly, applying the algorithm (3.1.1) in the case of σ is constant. The Nth-order phase transformation or the reference IF function estimate fωxN

is

3.1.2 The higher-order short time Fourier synchrosqueez-ing transform

Higher-order FSST

Definition 3.1.1. Given a signal x(τ ) = A(τ )ei2πφ(τ ) in L2(R) with A(τ ) and φ(τ ) are equal to their Lth-order and N -order respectively, the Taylor expansion for τ close to t (see[17]):

A(τ ) = elog(A(τ ))= expXL

Since (log(t))(k)(t) = 0 if L + 1 ≤ K ≤ N , we define the STFT for a signal Applying the derivative of STFT, we have:

∂tVxg(t, η) =

The Nth-order IF estimate follows:

Ren

where Pk,1(t, η) = Vxtk−1gVg (t,τ )

x(t,τ ) , for k = 2, . . . , N.

Moreover, we can put equation (3.1.5) in the form

wx(t, η) = [1, P2,1(t, η), . . . , PN,1(t, η)]

In the same way, we can present the previous cases to the algorithm 3.1.1 to provide the parameters r1, . . . , rN.

Definition 3.1.2. If s ∈ L2(R), the Nth-order local complex IF estimate or phase transformation wNx is defined by (see [17]):

wNx(t, η) =

Adaptive higher-order FSST

We recall the STFT of the signal x(t) denoted by

Vfxg(t, η) =

Taking the derivative by t

To simplify the calculation, we note the following expression

Thus, we have

1

Then,

The IF function is obtained by

φ(1)(t) = Re in the form of a scalar product previously solved by the algorithm (3.1.1):

wex(t, η) = [1, S2,1(t, η), . . . , SN,1(t, η)]

∂ηSj,j−1(t, η). In the same way, algorithm (3.1.1) solves the problem.

3.2 New FSST transform

A new phase transformation for the 2nd-order adaptive FSST was proposed in [26]. We consider the new second order FSST associated with STFT. For a signal x(t) defined by

x(t) = Aept+q2t2ei2π(ct+12rt2),

The STFT of x(t) ∈ L2(R) with a window function g(t) ∈ L2(R) is defined as

Vx(t, η) = Z

R

x(τ )g(τ − t)e−i2πη(τ −t)

dτ (3.2.1)

where t is the time variable and η are is the frequency variable.

As in previous cases, we have

tVx(t, η) = (p + qt + i2π(c + rt))Vx(t, η) + (q + i2πr)Vxg1(t, η).

Thus at (t, η) on which Vx(t, η) 6= 0

tVx(t, η)

Vxg1(t, η) = p + qt + i2π(c + rt) Vx(t, η)

Vxg1(t, η) + (q + i2πr).

Taking partial derivative ∂η, then we have

∂η

∂tVx(t, η) Vxg1(t, η)

= p + qt + i2π(c + rt)P0(t, η) (3.2.2)

where we use P0(t, η) to denote

P0(t, η) = ∂

∂η

 Vx(t, η) Vxg1(t, η)



⇒ p + qt + i2π(c + rt) = 1 P0(t, η)

∂η

 Vx(t, η) Vxg1(t, η)

 .

Thus,

c + rt = −p + qt

i2π + 1

i2πP0(t, η)

∂η

 Vx(t, η) Vxg1(t, η)

 .

Thus for a general x(t), we define a new phase transformation for the 2nd-order FSST, denoted by ωxN ew,2nd, as

ωxN ew,2nd(t, η) =

3.2.1 New higher-order FSST

Definition 3.2.1. Given a signal x(τ ) = A(τ )ei2πφ(τ ) in L2(R) with A(τ ) and φ(τ ) are equal to their Lth-order and N -order respectively, the Taylor expansion for τ close to t:

A(τ ) = elog(A(τ ))= expXL

Applying the derivative of STFT, we have:

tVxg(t, η) =

Taking the derivative by η for the equation (3.2.4), we have:

ηwx(t, η) =(log(A))0(t)

Therefore, if in addition, ∂η

Vxg(t,η)

where the function wN ewx is defined by

wxN ew(t, η) = 1

In addition, we can write equation (3.2.4) in the form

The new version for the Nth-order IF estimate is defined by the STFT:

<n

Moreover, we can put the equation (3.2.4) in the form:

wN ewx (t, η) = [1, W3,1(t, η), W4,1(t, η), . . . , WN,1(t, η)]

In the same way the previous cases used the algorithm (3.1.1) to provide the parameters r1, r3, r4, . . . , rN, we can denote k = 3, . . . , N by computing the partial derivatives:

y2(t, η) = ∂ηωN ewx (t, η)

ηW3,1(t, η) and Wk,2(t, η) = ∂ηWk,1(t, η)

ηW3,1(t, η) (3.2.7) which implies the following expression

y2(t, η) = [0, 1, W4,2(t, η), . . . , WN,2(t, η)]RTN −1. (3.2.8)

To find the jthequation, we do the same process iteratively. We define the

We group the N − 1 equations and get a good linear system:

Since the AN −1 is an upper triangular matrix with a nonzero diagonal, the solution of the linear system is given by

rN(t) = yN −1(t, η) Definition 3.2.2. Let x ∈ L2(R). The New version for Nth-order local complex IF estimate or phase transformation ωxN,N ew is defined by

ωxN,N ew(t, η) =

Chapter 4

Numerical simulation

4.1 Numerical Simulation

In this section, we present some experimental results for the new second order of the phase transformation ωx2nd. Let x(t) be a signal with two linear chirps:

x(t) = x1(t) + x2(t) = cos 2π(c1+ 1

2b1t)t + cos 2π(c2+1

2b2t)t, t ∈ [0, 1]

(4.1.1) where the reference frequencies are c1 = 12, c2 = 34, and the chip rates are b1 = 50, b2 = 64. Here, x(t) is sampled uniformly with N = 256 sample points.

We proceed with two representative signal types.

Example 1: Our first signal is a signal with three components, given by s(t) = s1(t) + s2(t) + s3(t),

where

s1(t) = cos(118π(t −12) + 100π(t −12)2)1[1

2,1]. s2(t) = cos(94πt + 110πt2+ 13 cos(4πt − pi2)).

s3(t) = cos(194πt + 112πt.2)1[0,3

4].

(4.1.2)

Figure 4.1: The signal s(t) and its components s1(t), s2(t) and s3(t)

We use the relative “root mean square error” (RMSE) to evaluate the separation performance, which is defined by

RM SE = 1 K

K

X

k=1

||zk− ˆzk||2

||zk|| ,

where ˆzkis the reconstruction result of zk, K is the number of components.

Here, we test some parameters of σ from 0.001 to 0.1. The best value for the reconstrucation signal and its components from FSST2 can be obtained by minimizing the RMSE for both approaches of FSST2. Then, σOld = σN ew = 0.023. We colculated the following results.

σ 0.015 0.020 0.023 0.025 0.030 RMSE for Old 0.1375 0.0860 0.0798 0.0843 0.0922 RMSE for New 0.1091 0.0720 0.0717 0.0782 0.0888

Table 4.1: Some differents values of σ and their RMSE.

Figure 4.2: RMSE for FSST2 Old and New with σ ∈ [0.001, 0.04]

As we can see, the best value for minimizing the value of RMSE is σ ≈ 0.023. This value shows the difference between the original IFs and the reconstrucation using the FSST2 old and FSST2 new.

Figure 4.3: The original IF of the signal s(t)

Figure 4.4: Difference of reconstructed IFs with original IFs by old 2nd-order and new 2nd-order FSST.

Figure 4.5: Difference for the reconstructed s1, s2, s3with original component s1(t), s2(t), s3(t) by old 2nd-order and new 2nd-order FSST.

Example 2: The second signal is a signal with two components, given by the signal s(t) = s1(t) + s2(t), defined by

s1(t) = cos(2π(12t + 25t2)) and s2(t) = cos(2π(34t + 32t2)). (4.1.3)

Figure 4.6: The signal s(t)

Figure 4.7: The components of the signal s(t) one by one s1(t)(left), and s2(t)(right)

Here, we test some parameters of σ from 0.001 to 0.1. The best value for the reconstrucation signal and its components from FSST2 can be obtained by minimizing the RMSE for both approaches of FSST2. Then, σOld = σN ew ≈ 0.05. We colculated the following results.

σ 0.04 0.042 0.045 0.047 0.050

RMSE for Old 0.0893 0.0897 0.0802 0.0827 0.0824 RMSE for New 0.0893 0.0893 0.0800 0.0823 0.0822

Table 4.2: Some Examples for differents parameter of σ.

Figure 4.8: RMSE for Old and New FSST2 with σ ∈ [0.001, 0.1]

As we can see, the best value for minimizing the value of RMSE is σ ≈ 0.045. This value demonstrates the difference between the original IFs and the reconstrucations using the FSST2 old and FSST2 new.

Figure 4.9: The original IF of the signal s(t)

Figure 4.10: Difference of reconstructed IFs with original IFs by old 2nd-order and new 2nd-2nd-order FSST.

Figure 4.11: The difference for the signal s(t)

Figure 4.12: Difference for the reconstructed s1(t), s2(t) with original com-ponent s1(t), s2(t) by old 2nd-order and new 2nd-order FSST.

Chapter 5

Analysis of Adaptive Shor-time Fourier Transform-based

Synchrosqueezing Transform

5.1 Analysis for new approach of FSST2

The Fourier transform of eiπφ00k(t)τ2g(τ ), which we denote by Gk(ξ), is defined by (refer to [19]):

Gk(ξ) = F

eiπ)φ00k(t)τ2g(τ ) ξ

= Z

R

eiπφ00k(t)τ2g(τ )e−i2πξτ

xk(t + τ ) = xk(t)ei2π(φ0k(t)τ +12φ00k(t)τ2)+ Ak(t + τ ) − Ak(t)ei2πφk(t+τ ) + xk(t)ei2π(φ0k(t)τ +12φ00k(t)τ2) ei2π φk(t+τ )−φk(t)−φ0k(t)τ −12φ00k(t)τ2



− 1.

Then, we have Vx(t, η) =

K

X

k=1

Z

R

xk(t + τ )g(τ )e−i2πητ

=

K

X

k=1

Z

R

xk(t)ei2π(φ0k(t)τ +12φ00k(t)τ2)g(τ )e−i2πητdτ + res0.

where

res0 = PK k=1

R

R{ Ak(t + τ ) − Ak(t)ei2πφk(t+τ )

+ xk(t)ei2π(φ0k(t)τ +12φ00k(t)τ2) ei2π φk(t+τ )−φk(t)−φ0k(t)τ −12φ00k(t)τ2



− 1}g(τ )e−i2πητdτ.(5.1.1)lll

|res0| ≤Y

0

(t), (5.1.2)

where

Y

0

(t) = Kε1I1+ π 3ε3I3

K

X

k=1

Ak(t).

We introduce more notations defined as follows

Gj,k(t, η) = Z

R

ei2π(φ0k(t)τ +12φ00k(t)τ2)τjg(τ )ei2πητdτ.

= F



eiπ(φ00k(t)τ2)τjg(τ )



η − φ0k(t).

We also note

Gk η − φ0k(t) = G0,k(t, η) Gj,k = 1

(−i2π)jG(j)k η − φ0k(t).

Lemma 5.1.1. [19] Let x(t) = A(t)ei2πφ(t) ∈ L2(R), we have

tVx(t, η) = i2πφ0k(t)Vx(t, η) + i2πφ00k(t)Vxg1(t, η) + Res1 (5.1.3)

where

Res1 = i2πBk(t, η) + i2πDk(t, η) + i2π(η − φ0k(t))res0− res00− i2πφ0k(t)res1.

Proof. Vxg0 is defined by (see [19]):

Vxg0 = i2π

K

X

`=1

x`(t) η − φ0`(t)G0,`(t, η) − i2π

K

X

`=1

x`(t)φ00`(t)G1,`(t, η) + res00.

Taking the derivatve by the time variable t, we obtain

tVx(t, η) = i2πηVx(t, η) − Vxg0(t, η). (5.1.4)

Using the derivative of the time variable (5.1.4), we have

tVx(t, η) − i2πφ0k(t)Vx(t, η) − i2πφ00k(t)Vxg1(t, η)

= i2π η − φ0k(t)Vx(t, η) − Vxg0(t, η) − i2πφ00k(t)Vxg1(t, η)

= i2π η − φ0k(t)Vx(t, η) − i2π

K

X

`=1

x`(t) η − φ0`(t)G0,`(t, η) − res00

+ i2π

K

X

`=1

x`(t)φ00`(t)G1,`(t, η) − i2πφ00k(t)Vxg1(t, η).

In addition, we define the expression of Vx(t, η) and Vxg1(t, η) in the form

Vx(t, η) =

K

X

`=1

x`(t)G0,`(t, η) + res0

Vxg1(t, η) =

K

X

`=1

x`(t)G1,`(t, η) + res1.

Then, we have

tVx(t, η) − i2πφ0k(t)Vx(t, η) − i2πφ00k(t)Vxg1(t, η)

We recall the definition of the new phase transformation for the 2nd-order STFT, which Vx(t, η) 6= 0.:

Then, the new second order complex IF is defined as follows

ωN ew,2nd,cx (t, η) = P1(t, η) i2πQ1(t, η)

where P1 and Q1 are defined by

P1(t, η) = ∂

∂η

∂tVx(t, η) Vxg1(t, η)



and Q1(t, η) = ∂

∂η

 Vx(t, η) Vxg1(t, η)



. (5.1.5) Lemma 5.1.2. For (t, η) such that Vxg1(t, η) 6= 0 and P1(t, η) 6= 0, as defined in the equation (5.1.5), we have

P1(t, η) = i2πφ0k(t)Q1(t, η) + Res3 (5.1.6)

and we note that Res2 = ∂ηRes1(see[19]), where

Res3 = Res2Vxg1(t, η) − ∂ηVxg1(t, η)Res1 Vxg1(t, η)2 . Proof. Using the result of lemma (5.1.1), we have

tVx(t, η) = i2πφ0k(t)Vx(t, η) + i2πφ00k(t)Vxg1(t, η) + Res1. (5.1.7)

Therefore, in the equation (5.1.7), therefore, thus

tVx(t, η)

Vxg1(t, η) = i2πφ0k(t) Vx(t, η)

Vxg1(t, η) + i2πφ00k(t) + Res1

Vxg1(t, η). (5.1.8) Taking the derivative by η, then yields

η

∂tVx(t, η) Vxg1(t, η)



= i2πφ0k(t)∂η

 Vx(t, η) Vxg1(t, η)

 + ∂η

 Res1

Vxg1(t, η)



. (5.1.9)

We replace in the equation (5.1.9) with the expression of the P1 and Q1:

P1(t, η) = i2πφ0k(t)Q1(t, η) + Res3. (5.1.10)

This completes the proof of lemma (5.1.2).

Theorem 5.1.3. Let x(t) ∈ Dε12 for small ε1, ε2 > 0. We have the follow-ing results:

(a) Suppose ε1 satisfies ε1 ≥ Π0(t) + τ0

K

X

k=1

ckAk(t), and for (t, η) with

|Vxg1(t, η)| > ε1. Then, there exists k ∈ {1, . . . , K} such that (t, η) ∈ Ok

(b) If (t, η) such that |Vx(t, η)| > ε1, ∂η

 Vx(t,η) Vxg1(t,η)



> ε2 and (t, η) ∈ Ok, then, we have

ωN ew,2nd,cx (t, η) − φ0k(t) = Res4

where

Res4 = Res2Vxg1(t, η) − ∂ηVxg1(t, η)Res1 i2π

ηVx(t, η)Vxg1(t, η) − ∂ηVxg1(t, η)Vx(t, η) . In addition, we have

N ew,2nd(t, η) − φ0k(t)| < Bd1 where

Bd1 = max

1≤k≤K sup

η∈Ok

n 1

2πε21ε2 |Res2||Vxg1(t, η)| + |Res1||∂ηVxg1(t, η)|o

(c) If ε1 satisfies the condition in part (a) and Bd112Lk(t), then

and with |Zt| represents the Lebesgue measure of the set Zt:

Zt:=n

which contradiction the assumption |Vxg1(t, η)| > ε1. Therefore, (a) holds.

Proof Part (b). Using the result of Lemma (5.1.2),

P1(t, η) = i2πφ0k(t)Q1(t, η) + Res3. (5.1.12)

Then we have

ωN ew,2nd,cx (t, η) − φ0k(t) = Res4

where

Res4 = Res3

i2πQ1(t, η).

= Vxg1(t, η)2

Res3 i2π



ηVx(t, η)Vxg1(t, η) − ∂ηVxg1(t, η)Vx(t, η)

 .

However, the formula of Res3 is defined in the lemma (5.1.2), and we obtain

Vxg1(t, η)2

Res3 = Res2Vxg1(t, η) − ∂ηVxg1(t, η)Res1.

Then,

Res4 = Res2Vxg1(t, η) − ∂ηVxg1(t, η)Res1

i2π

ηVx(t, η)Vxg1(t, η) − ∂ηVxg1(t, η)Vx(t, η) .

Next, we have

|Res4| = Proof Part (c). First, we have the following result from [7] on p.254

lim

Thus, η ∈ Xt. This which implies the first inclusion Yt ⊂ Xt.

Using theorem (2.6.4) part(a), if η ∈ Xt,

then |Vxg1(t, η)| > ε1 and there exists l ∈1, 2, . . . , K such that (t, η) ∈ Ol. If l 6= k, then

x,εN ew,2nd12 (t, η) − φ0k(t)| ≥ |φ0k(t) − φ0l(t)| − |φ0l(t) − ωx,εN ew,2nd12 (t, η)|.

We use the following inequalities

0k(t) − φ0l(t)| > Lk and |φ0l(t) − ωx,εN ew,2nd

12 (t, η)| < Bd1 ≤ ε3. Therefore,

ωx,εN ew,2nd12 (t, η) − φ0k(t)

> Lk(t) − ε3 ≥ ε3, and contradicts the assumption η ∈ Xt.

Therefore, l = k and η ∈ Yt, implying the second inclusion Xt = Yt. We recall

Zt:=n

η : (t, η) ∈ Ok,

Vxg1(t, η) > ε1,

η Vx(t, η) Vxg1(t, η)



≤ ε2o ,

The fact that Xt = Yt and Yt∩ Zt= φ, with the equation (5.1.13), imply that

Z

Using this equalition

Z For the first term, we get the following results:

T erm1 =

substituting the parameters estimated above we get We group all parameter estimates, we find the following results

|

Hence, we have

Then we have the result

lim

λ→0

1 g(0)

Z

|ξ−φ0k(t)|<ε3

Radp,2nd,λx,ε12 (t, ξ)dξ − xk(t)

≤ Bd2,

This completes proof of Theorem. 2

Chapter 6

Conclusion and future work

In this study, we introduced a generalization of the STFT-based SST (FSST) with time-varying, CWT-based SST (WSST), the WSST with time-varying, and the FSST with a new phase transformation by using higher order ampli-tude and phase approximations. This generalization allows us to better assess a wide variety of multicomponent signals containing very strongly modulated AM-FM modes.

We also studyed the theoretical analysis of the 2nd-order FSST with a new phase transformation. The new phase transformation is much simpler than the convectional one. The new FSST performance in IF estimation and component recovery is comparable with that of the conventional 2nd-order FSST. In some cses, the new FSST perfermed even better than its conven-tional counterpart.

Since the result showed a better concentration and reconstruction for a wider variety of AM-FM modes, we will continue working on the adaptive FSST with a new phase transformation by using higher order approximations both for the amplitude and phase.

Bibliography

Bibliography

[1] L. Cohen, Time-frequency Analysis, Prentice Hall, New Jersey, (1995).

[2] F. Hlawatsch and G.F. Boudreaux-Bartels, Linear and quadratic time-frequency signal representations,IEEE signal processing magazine., 9.2(1992)21–67.

[3] B. Boashash, Time-frequency signal analysis and processing: a compre-hensive reference, Academic Press., (2015).

[4] P. Flandrin, Time-frequency/Time-scale analysis, Wavelet Analysis and its Applications., (1999).

[5] S. Mallat, A wavelet tour of signal processing, Elsevier., (1999).

[6] I. Daubechies, S. Maes,A nonliner sequeezing of the continuous wavelet transform based on auditory never models. Wavelets in medicine and Biology., (1996) 527-546.

[7] I. Daubechies, J. Lu, and H .-T. Wu, Synechrosqueezed wavelet trans-fonns: An empirical mode decomposition-like tool. Applied and compu-tational harmonic analysis., 30.2 (2011) 243-261.

[8] S. Meignen, T. Oberlin. and S. McLaughlin, A new algorithm for mul-ticomponent signals analysis based on synchrosqueezing: With an ap-plication to signal sampling and denoising. IEEE transactions on Signal Processing., 60.11. (2012 )5787—5798.

[9] G. Thakur and H.-T. Wu, Synechrosqueezing-based recovery of instanta-neous frequency from nonuniform samples, SIAM Journal on Mathemat-ical Analysis., 43.5 (2011) 2078-2095.

[10] T.Oberlin, S. Meignen, and V. Perrier, Second-order synchrosqueezing transform or invertible reassignment? Towards ideal time-frequency rep-resentations, IEEE Transactions on Signal Processing., 63.5(2015) 1335-1344.

[11] T. Oberlin and S. Meignen, The second-order wavelet synchrosqueezing transform, 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE., (2017).

[12] H.-T. Wu, Adaptive analysis of complex data sets, Phd. dissertation, Princeton University, Princeton, NJ., (2012)

[13] D. Iatsenko, P.-V. E. McClintock and A. Stefanovska, Linear and syn-chrosqueezed time-frequency representations revisited: Overview, stan-dards of use, resolution, reconstruction, concentration, and algorithms, Digital Signal Processing., 42.7 (2015) 1-26.

[14] L.Li, H.Cai, Q. Jiang, and H. Ji, Adaptive Synchrosqueezing transform with a tim-varying parameter for non-stationary signal separation. Ap-plied and Computational Harmonic Analysis., (2019).

[15] L.Li, H.Cai, H. Han, Q. Jiang, and H. Ji, Adaptive short-time Fourier transform and synchrosqueezing transform for non-stationary signal sep-aration, Signal Processing., 166 (2020)107231.

[16] Y.-L. Sheu, L.-Y. Hsu, P.-T. Chou, and H.-T. Wu, Entropy-based time-varying window width selection for nonlinear-type time-frequency anal-ysis, International Journal of Data Science and Analytics., 3.4 (2017) 231-245.

[17] D.-H. Pham and S. Meignen, High-order synchrosqueezing transform for multicomponent signals analysis - with an application to gravitational-wave signal, IEEE Transactions on Signal Processing., 65.12 (2017) 3168-3178.

[18] R. Behera, S. Meignen, and T. Oberlin, Theoretical analysis of the second-order synchrosqueezing transform, Applied and computational harmonic Analysis., 45.2 (2018) 379–404.

[19] H. Cai, Q. Jiang, L. Li, and Bruce W. Suter, Analysis of adaptive short-time fourier transform-based synchrosqueezing transform, arXiv preprint arXiv.,1812.11033 (2018).

[20] I. Daubechies, Ten Lectures on Wavelets, CBMF Conference Series in Applied Mathematics, Vol. 61 (SIAM, Philadelphia, PA, 1992).

[21] C. L, Herry. M. Frasch,A.J. Seely,and H.-T. Wu, Heart beat classification from single-lead ECG using the synchrosqueezing transform. Physiologi-cal Measurement., 38.2(2017).

[22] C.K. Chui, An Introduction to wavelets, Academic Press., (1992).

[23] Y. Meyer, Wavelets and operators, Volume 1, Cambridge university Press., (1993).

[24] C.K. Chui and Q.T. Jiang, Applied mathematics. Data Compression, Spectral Methods, Fourier Analysis, Wavelets and Applications., (2013).

[25] T. Oberlin, S. Meignen, and V. Perrier, The Fourier-based syn-chrosqueezing transform, 2014 IEEE international conference on acous-tics, speech and signal processing (ICASSP), IEEE., (2014).

[26] Q. Jiang, The second-order synchrosqueezing transform with a new phase transformation, preprint, (2018).

[27] E. Sejdic, I. Djurovic and J. Jiang, Time-frequency feature representa-tion using enery concentrarepresenta-tion: An overview of recent adavances, Digital signal processing., 19.1 (2009) 153–183.

[28] A. Berrian and S. Naoki, Adaptive synchrosqueezing based on a quilted short-time Fourier transform. In Wavelets and Sparsity XVII, Vol. 10394.

International Society for Optics and Photonics., (2017).

[29] L.Li, H.Cai, Z. Wang, Q. Tang, and H. Ji, The Higher-order wavelet synchrosqueezing transform, preprint, (2018).

[30] F. Auger and P. Flandrin, ”Improving the readability of time-frequency and time-scale representations by the reassignment method, IEEE Trans-actions on signal processing., 43.5(1995):1068—1089.

[31] E.Chassande-Mottin, F. Auger, and P. Flandrin, Time-frequency/time-scale reassignment, Wavelets and signal processing, Birkh¨auser, Boston, MA., (2003). 233-267.

[32] N.E. Huang, Z. Shen, S.R. Long, M.L. Wu, H.H. Shih, Q. Zheng, N.C.

Yen, C.C. Tung, and H.H. Liu, The empirical mode decomposition and Hilbert spectrum for nonlinear and non-stationary time series analysis.

Proceeding of the Royal Society of London, Series A:mathematical, phys-ical and engineering sciences., 454.1971(1998) 903-995.

[33] P. Flandrin, G. Rilling, and P. Goncalves, Empirical mode decomposi-tion as a filter bank, IEEE signal Processing Letters., 11.2 (2004) 112-114.

[34] Y. Xu, B. Liu, J. Liu, and S. Riemenschneider, Two-dimensional em-pirical mode decomposition by finite elements, Proceedings of the Royal Society London A: Mathematical, Physical and Engineering Sciences., 462.2074 (2006): 3081-3096.

[35] G. Rilling and P. Flandrin, ”One or two frequencies? The empirical mode decomposition answers, IEEE transactions on signal processing., 56.1 (2007) 85-95.

[36] Z. Wu and N.E. Huang, “Ensemble empirical mode decomposition: a noise-assisted data analysis method. Advances in adaptive data analysis., 1.01 (2009) 1-41.

[37] A. Cicone, J.F. Liu, and H.M. Zhou, “Adaptive local iterative filtering for signal decomposition and instantaneous frequency analysis, Applied Computational Harmonic Analysis., 41.2(2016) 384–411.

[38] L. Cohen, Time-frequency distributions-a review,Proceedings of the IEEE., 77.7 (1989) 941-981.

[39] S. Qian and D. Chen, Joint time-frequency analysis, IEEE Signal Pro-cessing Magazine., 16.2(1999) 52–67.

[40] H.-T. Wu, P. Flandrin, and I. Daubechies, One or two frequencies?

The synchrosqueezing answers, Advances in Adaptive Data Analysis., 3.4 (2011) 29-39.

[41] L. Li and H. Ji, Signal feature extraction based on improved EMD method. Measurement., 42.5 (2009) 796-803.

[42] L.Lin, Y. Wang, and H.M. Zhou, “Iterative filtering as an alternative algorithm for empirical mode decomposition, Advances in Adaptive Data Analysis., 1.04 (2009) 543-560.

[43] S. Meignen, D.-H. Pham, and S. McLaughlin, On demodulation, ridge detection and synchrosqueezing for multicomponent signals, IEEE Trans-actions on Signal Processing., 65.8 (2017) 2093-2103.

[44] C. Li and M. Liang, A generalized synchrosqueezing transform for en-hancing signal time-frequency representation, Signal Processing, 92.9 (2012) 2264-2274.

[45] H.Z. Yang, Synchrosqueezed wave packet transforms and diffeomor-phism based spectral analysis for 1D general mode decompositions, Ap-plied and Computational Harmonic Analysis., 39.1 (2015) 33-66.

[46] A.J. Berrian, The Chirped quilted synchrosqueezing transform and its application to bioacoustic signal analysis, Phd. dissertation, University of California at Davis., (2018).

[47] C.K. Chui, Y.-T. Lin, and H.-T. Wu, Real-time dynamics acquisition from irregular samples-with application to anesthesia evaluation, Analysis and Applications., 14.04 (2016) 537-590.

[48] C.K. Chui and H.N. Mhaskar, Signal decomposition and analysis via ex-traction of frequencies, Applied and Computational Harmonic Analysis., 40.1 (2016) 97-136.

[49] C.K. Chui and M.D. van der Walt, Signal analysis via instantaneous frequency estimation of signal components, GEM-International Journal on Geomathematics., 6.1 (2015) 1-42.

[50] Q.T. Jiang and B.W. Suter, Instantaneous frequency estimation based on synchrosqueezing wavelet transform, Signal Processing., 138 (2017) 167-181.

In document Vista de Exemplar complet (página 55-60)